Best DSL MTU not 1500?

[cahaddras]

The advice on this forum and the speedguide analyser is that the optimum MTU for a DSL connection is 1500. Surely this is wrong? For a PPPoA connection, if the speed is limited only by the modem speed, then an MTU value of 1478 is usually the best as this is the maximum value that avoids wasted padding bytes in the fixed size ATM cells that carry data over the link. For the same reason, an MTU value of 1454 is usually best on a PPPoE connection, although PPPoA is better if your ISP supports it.

[mccoffee]

Most of the time people have pppoe since atm is very expesive. the mtu should 1480 for xp, or 1492. Now for some odd reason though like bell south like the 1454 why I don't know like you said probably less padding. With pppoa I was told and read it's 1500 for mtu. However it's not really the modem that caps the speed it's more the network that does that. Actully it's funny a dsl modem isn't really a modem it's mutiplexer.

Each provider's network is different best bet to ping to find mtu for your network.

[cahaddras]

All DSL connections use ATM - it's built in to the protocol. PPPoE is less efficient than PPPoA because PPPoE is really PPP over Ethernet over ATM, and the extra layer wastes an extra 24 bytes protocol overhead in every packet. That's why the optimum MTU for PPPoE is 24 bytes less (1478 - 24 = 1454) than that for PPPoA - it's the same maximum packet size over ATM, but allowing for the extra header. In each case you actually end up with 1488 bytes carried over ATM for the maximum case, and the reason that's optimum is that ATM carries data in fixed size cells, 48 data bytes per cell, and 1488 is an exact multiple of 48.

As you say though, if it's not the modem that's capping the spead then fitting exactly into ATM cells is not important. Then the protocol maximum MTU is best - 1500 for PPPoA and 1492 for PPPoE. However in the UK at least it's quite common nowadays to have the modem limiting, at 8Mbps for ADSL or higher for ADSL2+, and PPPoA is used widely for best efficiency. In this case tuning the MTU down from 1500 to 1478 gives approximately a 2% improvement in throughput.

[mccoffee]

From what I read and studied though say for example you had the mtu set 1500 on the pppoe connection that those exra bits will get dropped anway since the network's mtu is 1492.


I had see where the some prvodiers such as sbc in the states now it's ATT not sbc used mtu 1454 on default for there pppoe software ethernet change to 1492 booming speed .

In bellsouths case i seen it work both ways either or 1454 or 1492. Anohter odd thing though some modems for dsl have a router mode and a bridge mode depending on mode the modem is router/bridge the mtu changes on some models for defautl mtu why not sure it's odd how some networks uses different values.

I noticed some vdsl2 providers in australlia used the 1454 as well.

I think I admi i have do some reading again if you did set mtu on pppoa for 1500 what ever the mtu on the network is that padding gets droped with out interuption or thoughput slowdown or it should though I could see how it could slow it done it's not suppose. I could be wrong.

[Andrzej]

it still returns as 'never ending story'

BTW in ADSL with PPPoA
every AMT cell size is 53 = 48(data)+5(ATMheader)
so in theory the best MTU for PPPoA should be find as follow:
MTUvalue = multiple 53 (ATM cell size)
but IMVHO it can be significant for connections bigger than 10MB

excellent explanation of ATM cell size and data transfer
and 'all headers' can be find in
http://www.geocities.jp/hamdah350221...i/mtu_1gif.gif link

OTOH in PPPoA total sum of 'headers' included to every it MTU
(see above link)
as a rule is bigger than in PPPoE
so PPPoA-MTU total data transfer efficiency is lower
(e.g. 86% vs 97%).

BTW lower MTU than 1492 can be find in:
Australian, British, Germans and Japans providers

[cahaddras]

Your diagram shows PPPoE, not PPPoA, but in any case you've misinterpreted the diagram on two counts: there are additional headers to take into account over the MTU before splitting into ATM cells, and as the cells hold 48 bytes of data, the packet is fragmented into 48 byte cells after which ATM adds a 5 byte header to each. The diagram actually shows a DSL variant known as PPP over LLC encapsulated bridged ethernet. This is pretty much the worst case in terms of efficiency, but in this case a 1448 byte MTU is optimum. This MTU size includes the 20 byte IP and 20 byte TCP header, so it actually carries 1408 bytes of data. Before transmission, a further 32 bytes of header are added, consisting of PPP(2), PPPoE(6), ethernet MAC (14), PAD(2), LLC+SNAP (8), plus another 8 bytes of CCPS-PDU trailer, goving a total of 1488 bytes. This is then fragmented exactly into 31 cells of 48 bytes, each of which then gets a further 5 byte cell header giving a total of 1643 bytes transmitted at whatever speed the modem has synchronised at. So the overall efficiency in this case is 1408 data bytes requiring 1643 line bytes, which is 85.7%.

In reality an alternative called VC based multiplexing is often used over DSL instead of LLC encapsulation, saving the largely redundant 8 byte LLC + SNAP header (see http://www.ietf.org/rfc/rfc1483.txt for details). Then the optimum MTU increases to 1456, carrying 1416 bytes of data using 1643 bytes on the line, giving an efficiency of 86.2% for VC multiplexed PPPoE.

PPPoA with VC multiplexing is the most efficient configuration for DSL. This further misses out the 6 PPPoE, 14 ethernet MAC and 2 PAD bytes. So you end up with just the 2 byte PPP header plus the 8 byte CPCS-PDU trailer and ATM cell headers (see http://www.ietf.org/rfc/rfc2364.txt). Then the optimum MTU increases to 1478 bytes, carrying 1438 bytes TCP data over 1643 line bytes, and giving an efficiency of 87.5%.

[Philip]

Cahaddras, thanks for the input. Note that the Analyzer does not distinguish between DSL/Cable or other types of connections, it simply interprets TCP/IP headers. Therefore, the recommended MTU is generic, the maximum possible to reduce header overhead.

The assumption behind 1492 recommendation for PPPoE is that the line is limited at the modem at your end, and it does not take into account ATM frames. The MS default 1480 MTU is better tuned for complete ATM frames, however, as you've already pointed out, it all depends on the exact implementation.

It would be interesting to systematically work out/illustrate the different possibilities, write them out so everyone understands them and put them up for discussion...

Still, tweaking depends on where the bottleneck of the connection is, ATM overhead might be more of the ISP's concern.

[pgogborn]

Originally Posted by cahaddras

So you end up with just the 2 byte PPP header plus the 8 byte CPCS-PDU trailer and ATM cell headers (see http://www.ietf.org/rfc/rfc2364.txt). Then the optimum MTU increases to 1478 bytes, carrying 1438 bytes TCP data over 1643 line bytes, and giving an efficiency of 87.5%.

cahaddras, I found this very interesting.

A few days ago I discovered my ISP supplier in the UK, which uses PPPoA with VC multiplexing, was using a middlebox to change my MTU from 1500 to 1478.

In surfing the net to find a logical explanation for why they did it, I came across your post. Thank you.